Qwen3-Omni

2025-09-22 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba Qwen releases Qwen3-Omni-30B-A3B, an open-weights omni-modal MoE model (Instruct, Thinking and Captioner) that processes text, image, audio and video and responds in text or streaming speech.

### Context Qwen2.5-Omni (March 2025) had opened the category of end-to-end omni models. Qwen3-Omni takes the idea up on the Qwen3 MoE base and is compared with Gemini 2.5 Pro, Gemini 2.5 Flash, GPT-4o and GPT-4o-Transcribe on audio and video. ### What the model brings - Thinker-Talker MoE architecture with pre-training of an audio encoder, "AuT", and a multi-codebook design to reduce speech latency. - Vendor claim: state of the art on 22 of 36 audio and audio-visual benchmarks, and open-source state of the art on 32 of 36; speech recognition, audio understanding and voice conversation comparable to Gemini 2.5 Pro. - Multilingual: 119 text languages, 19 speech input languages, 10 speech output languages. - No regression on text and image compared with Qwen models of the same size, thanks to mixed multimodal pre-training that starts with text. - Behavior control through the system prompt, and an open Captioner model for detailed audio description. ### Architecture and specifications Model named 30B-A3B (about 35 billion parameters in total on Hugging Face, Thinker and Talker included). Variants: Instruct (Thinker and Talker), Thinking and Captioner. Apache 2.0 license (declared on the Hugging Face card). Theoretical minimum GPU memory in BF16 for Instruct: 78.85 GB (15 s video) to 144.81 GB (120 s); for Thinking: 68.74 GB to 131.65 GB. Disabling the Talker saves about 10 GB. Knowledge cutoff: not disclosed. ### Cost and access | Item | Value | |---|---| | Weights | Free, Apache 2.0 | | qwen3-omni-flash API (current pricing consulted), text | $0.43 input / $1.66 output per million tokens | | qwen3-omni-flash API, audio input | $3.81 per million tokens (audio output $3.06) | | qwen3-omni-flash API, image or video input | $0.78 per million tokens (output $15.11) | ### Benchmarks Text, Instruct, Hugging Face card (vendor measurements). | Benchmark | GPT-4o-0327 | Qwen3-Omni-30B-A3B-Instruct | |---|---|---| | MMLU-Redux | 91.3 | 86.6 | | GPQA | 66.9 | 69.6 | | AIME25 | 26.7 | 65.0 | | MultiPL-E | 82.7 | 81.4 | | IFEval | 83.9 | 81.0 | LibriSpeech speech recognition (word error rate, test clean then other, lower is better), same card: GPT-4o-Transcribe 1.39 then 3.75; Gemini 2.5 Pro 2.89 then 3.56; Qwen3-Omni-30B-A3B-Instruct 1.22 then 2.48. ### Versus closed models On LibriSpeech speech recognition, the model posts the lowest error rates in the row (1.22 then 2.48), ahead of GPT-4o-Transcribe and Gemini 2.5 Pro. On text, it is ahead of GPT-4o-0327 on GPQA (69.6 versus 66.9) and especially AIME25 (65.0 versus 26.7), and behind on MMLU-Redux (86.6 versus 91.3) and IFEval. The vendor claims parity with Gemini 2.5 Pro on speech recognition, audio understanding and voice conversation, as well as outperforming GPT-4o on some audio tests, without detailing all 36 tests here. The comparison uses closed models from the same period, but they were chosen and measured by the vendor. Self-hosting requires at least about 80 to 145 GB of GPU memory depending on video length. ### Key takeaways Qwen3-Omni moves open omni models to an MoE architecture with streaming speech and broad language coverage. The strongest results are on audio, where the vendor claims parity with the best closed models.

Tags : LLM, Open Source, Alibaba, Qwen, MoE, Multimodal, Audio, Apache 2.0